A multi-model ensemble approach for reservoir dissolved oxygen forecasting based on feature screening and machine learning

• A highly accurate and robust DO prediction model was developed by exploring the combination of feature selection and ensemble learning. • The effects of different times and sample sizes on the screening of MIC features were thoroughly considered to more scientifically use MIC to screen for key drivers affecting DO. • The ensemble-RF model was able to capture and simulate the trend of DO well during the model construction period, both during the period of steady change and during the period of sudden hypoxia. Dissolved oxygen (DO) concentration in aquatic systems plays a vital role in water aquaculture. An innovative approach that combines feature selection and ensemble learning to predict DO in aquatic ecosystems was proposed. Feature selection was first performed using Maximum Information Coefficient (MIC). Five machine learning algorithms were then employed to construct five hybrid-MIC models, including K-Nearest Neighbors (KNN), Backpropagation (BP) Neural Network, Long Short-Term Memory (LSTM), Kernel Ridge Regression (KRR), and Support Vector Regression (SVR). Finally, an ensemble-RF prediction model was built using Random Forests(RF). The main findings are as follows: (1) The MIC technique can effectively identify the key factors influencing DO. (2) The MIC significantly improves model performance. (3) The hybrid-MIC model was further improved by the ensemble-RF model, the average R 2 and NSE were both as high as 0.99, and the average MAE and RMSE were decreased by 72 % and 64 %, respectively.